Zapier Agents Is Now a Zap Step: What to Change

Zapier has folded its standalone Zapier Agents product into a single AI step inside the Zap editor, called AI by Zapier. Zapier's blog post announcing it is dated October 6, 2026. If you built agents on the old product, they are being converted into Zaps. If you never used Agents, the practical news is that an AI step with tools, per-tool approval and a task cap now sits next to Zapier's normal triggers, filters and paths.

For a business, this is less a new product than a change in where the risk lives. An agent used to be its own island. Now it is one step in a flow you can fence in on both sides.

Our sources: Zapier's announcement post, its help articles on migrating from Agents and adding tools to an AI by Zapier step, and Make's 2026 release notes, all read on October 16, 2026. We have not built a migrated agent ourselves for this post.

If you already run Zaps and want AI to make one judgment call inside them, the AI step is a good fit, and the per-tool approval toggle is the feature worth using. If you were running free-roaming agents with write access, migrate carefully, because one default can catch you out: Zapier's migration guide says tools run without pausing for confirmation unless you switch on approval for each one.

Zapier describes an agent as four things: a prompt, a set of tools, a trigger and permission to act. Those now live inside one step:

• Agent instructions become the prompt field.
• Tools become a tools picker inside the step.
• The trigger is the Zap trigger, so any app Zapier supports can start a run, not only a handful.
• A chat or on-demand agent becomes a webhook trigger.
• One agent calling another becomes a Sub-Zap trigger on a child Zap.
• Run history moves to Zap History, next to your other automations.

Zapier lists several things you can now do. You can mix AI steps with ordinary steps, so the AI picks the record and a normal action sends the confirmation email in the same wording every time. You can require approval per tool, so reading a spreadsheet runs freely while messaging a client waits for a person. You can define output fields, such as a category or a subject line, and get clean values to map into later steps. And each run shows the tools called, inputs, outputs, model tier and tasks used.

Zapier says AI steps bill as tasks, with default models free and other tiers carrying multipliers, so one AI step can cost 1, 3 or 5 tasks. You can set a task limit per run. Zapier says a run that passes 75 tasks pauses for a person instead of looping, and that this limit will become adjustable. Zapier's help article adds that tool use is dynamic, so the count only shows up in Zap History after the run.

That makes the AI step easy to leave running and hard to forecast. Check the first two weeks of Zap History before you trust a monthly number.

From Zapier's migration guide: eligible agents are converted automatically. Each becomes a Zap with a native trigger and one AI by Zapier step holding the original prompt, tools and reasoning. Migration does not split tool calls into separate steps, so the tools still run inside the AI step unless you move them out by hand. Turn off the original agent after you publish the converted Zap, or you can end up with two things doing the same job. Zapier also says knowledge sources are not yet supported for HubSpot, Asana, Zendesk, Zoom and Jira, and that an AI by Zapier step cannot be used as a tool inside another AI step.

Make is the closest comparison, and its release notes show a different emphasis. The October 9 entry covers AI Agent fallback connections, prompt caching and MCP tools for audit logs. A September 25 entry adds MCP tools for webhook queue management. A September 9 entry announces an official Make plugin for ChatGPT. The notes also carry plain maintenance warnings: Metricool's Get network metrics module stops returning 30 deprecated metrics from October 5, and an October 12 entry covers a Wealthbox OAuth 2.0 transition. We only have Make's headlines for these, not the details, so read the full entries before acting on one.

The pattern across both platforms is the same. The AI is getting easier to add. The boring parts, fallbacks, audit logs, deprecations and approvals, are where the work is.

These are designs, not client results.

1. Lead intake triage. A form submission starts the Zap. The AI step classifies the lead and drafts a reply from your own FAQ. A normal step writes the record to your CRM. The email goes out only after approval, set per tool. Everything the AI touches is read or draft. This pairs naturally with CRM hygiene work.

2. Social comment and message triage for managed social. A new comment or mention on your company channels starts the Zap. The AI step sorts it into question, complaint, sales interest or ignore, and drafts a reply into a review sheet. A person posts. We would not let an AI step publish replies on its own, partly for brand safety and partly because platform rules on automation differ and change. This is the kind of thing our managed social work already does by hand.

3. Invoice or form data cleanup. The AI step reads a messy attachment and returns defined fields. A plain step checks the totals and writes the row. If the check fails, the item goes to a person.

4. Weekly report summaries. The AI summarizes a spreadsheet. A deterministic step formats and sends it. Low risk, easy to test.

1. 1Turn on Require approval before running for every tool that sends, writes or deletes. Leave it off only for pure lookups.
2. 2Put a filter in front of the AI step so it only runs on records that need judgment. That keeps task counts down.
3. 3Define output fields instead of reading free text downstream.
4. 4Pick the model tier on purpose. Test the default first.
5. 5Write five to ten real test inputs, including ugly ones, and run them in Preview before publishing. See testing AI agents before production.
6. 6Name an owner for the Zap. Provider changes break automations, which is the real cost in automation maintenance.

• Default permissions after migration. No approval until you add it.
• Costs that vary by run. Check Zap History.
• Over-broad tools. Give the step only the apps it needs. See permissions and human approval.
• Knowledge source gaps. If your data lives in HubSpot or Zendesk, check the support list first.
• Double runs. Disable the old agent after migrating.

If the logic never varies, use a plain Zap with no AI step. It is cheaper and easier to explain. If volume is high and each item is low value, a multiplier on every run adds up. If you need version control, tests in code, or a self-hosted setup, a no-code step is the wrong tool, and our comparison of no-code and custom automation lays out where that line sits. And if the real problem is that the process itself is unclear, no AI step will fix it. See AI agents versus automations and the first automation mistake.

OrchestriAI does not resell Zapier or Make. We help decide which jobs deserve an AI step and which should stay plain. AI workflow automation covers the build, whether in a no-code platform or code. AI agent systems define tools and approvals. Systems integration handles the API connections underneath. If you are not sure which of your recurring tasks to start with, the free automation audit is built for that question.

OrchestriAI is an independent implementation provider. Product names are used for identification only. OrchestriAI is not affiliated with, endorsed by, or sponsored by Zapier or Make. This article summarizes Zapier's announcement of October 6, 2026, its help articles, and Make's release notes, all checked on October 16, 2026. Pricing, limits and supported apps can change, so confirm current details with each vendor before you move a live automation. This is not a security certification.